Anthropic's approach to safeguarding against Claude errors
Anthropic published a safety report on AI agent protections in claude.ai, Claude Code, and Claude Cowork. The key takeaway: robust environment-level isolation trumps attempts to make the model itself perfectly safe. The probabilistic nature of AI doesn't guarantee crash-free behavior.
Each product uses a distinct architecture. Claude.ai runs code in gVisor containers on Anthropic's servers. Claude Code uses a local sandbox based on macOS and Linux mechanisms. Claude Cowork relies on a full-fledged virtual machine that hides the host system from the agent.
The engineers also acknowledged past mistakes, such as attackers bypassing restrictions via allowed domains like api.anthropic.com. To prevent leaks, a custom proxy server was embedded inside the virtual machine.
Related stories
- One git call let a repo escape the Claude Code sandbox
- Stolen session keys are draining Claude Max accounts
- Breaking Claude Code Opus 5 Auto Mode
- Claude Opus broke four SAML libraries in a month
- Hidden China tracker pulled from Claude Code after exposé
- Auto mode let Claude Code run downloaded code in 6 of 10 tests
Comments
No comments yet. Be the first.
Join the conversation
Sign in with Google to leave a comment. Your name and avatar come from your Google profile, and the comment appears after moderation.